File size: 33,691 Bytes
b1e47cf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 | """
attribution_tracer.py
Core implementation of the Attribution Tracing module for the glyphs framework.
This module maps token-to-token attribution flows, tracks query-key alignment,
and visualizes attention patterns to reveal latent semantic structures.
"""
import logging
import time
import numpy as np
from typing import Dict, List, Optional, Tuple, Union, Any, Set
from dataclasses import dataclass, field
import json
import hashlib
from pathlib import Path
import matplotlib.pyplot as plt
import networkx as nx
from enum import Enum
from ..models.adapter import ModelAdapter
from ..utils.visualization_utils import VisualizationEngine
# Configure attribution-aware logging
logger = logging.getLogger("glyphs.attribution_tracer")
logger.setLevel(logging.INFO)
class AttributionType(Enum):
"""Types of attribution that can be traced."""
DIRECT = "direct" # Direct attribution between tokens
INDIRECT = "indirect" # Indirect attribution through intermediate tokens
RESIDUAL = "residual" # Attribution through residual connections
MULTIHEAD = "multihead" # Attribution through multiple attention heads
NULL = "null" # No clear attribution path
COMPOSITE = "composite" # Mixture of multiple attribution types
RECURSIVE = "recursive" # Self-referential attribution path
EMERGENT = "emergent" # Emerged from collective behavior, not individual tokens
@dataclass
class AttributionLink:
"""A link in an attribution chain between source and target tokens."""
source_idx: int # Index of source token
target_idx: int # Index of target token
attribution_type: AttributionType # Type of attribution
strength: float # Strength of attribution (0.0-1.0)
attention_heads: List[int] = field(default_factory=list) # Contributing attention heads
layers: List[int] = field(default_factory=list) # Contributing layers
intermediate_tokens: List[int] = field(default_factory=list) # Intermediate tokens in indirect attribution
residue: Optional[Dict[str, Any]] = None # Symbolic residue if attribution is weak/null
@dataclass
class AttributionMap:
"""Complete map of attribution across a sequence."""
prompt_tokens: List[str] # Tokenized prompt
output_tokens: List[str] # Tokenized output
links: List[AttributionLink] # Attribution links
token_salience: Dict[int, float] = field(default_factory=dict) # Salience of each token
attribution_gaps: List[Tuple[int, int]] = field(default_factory=list) # Gaps in attribution
collapsed_regions: List[Tuple[int, int]] = field(default_factory=list) # Regions with attribution collapse
uncertainty: Dict[int, float] = field(default_factory=dict) # Uncertainty in attribution
metadata: Dict[str, Any] = field(default_factory=dict) # Additional metadata
@dataclass
class ForkPath:
"""A fork in the attribution path, representing alternative attributions."""
id: str # Fork path ID
description: str # Description of the fork
links: List[AttributionLink] # Attribution links in this fork
confidence: float # Confidence in this fork (0.0-1.0)
conflict_points: List[int] = field(default_factory=list) # Tokens with conflicting attribution
residue: Optional[Dict[str, Any]] = None # Symbolic residue in the fork
@dataclass
class AttentionHead:
"""Representation of an attention head's behavior."""
layer: int # Layer containing this head
head: int # Head index
pattern_type: str # Type of attention pattern
focus_tokens: List[int] # Tokens this head focuses on
strength: float # Overall attention strength
function: Optional[str] = None # Inferred function of this head
attribution_role: Optional[str] = None # Role in attribution process
class AttributionTracer:
"""
Core attribution tracing system for the glyphs framework.
This class implements attribution tracing between tokens, mapping
how information flows through transformer architectures from inputs
to outputs. It provides insights into the causal relationships
between tokens and the formation of semantic structures.
"""
def __init__(
self,
model: ModelAdapter,
config: Optional[Dict[str, Any]] = None,
visualizer: Optional[VisualizationEngine] = None
):
"""
Initialize the attribution tracer.
Parameters:
-----------
model : ModelAdapter
Model adapter for the target model
config : Optional[Dict[str, Any]]
Configuration parameters for the tracer
visualizer : Optional[VisualizationEngine]
Visualization engine for attribution visualization
"""
self.model = model
self.config = config or {}
self.visualizer = visualizer
# Configure tracer parameters
self.trace_depth = self.config.get("trace_depth", 5)
self.min_attribution_strength = self.config.get("min_attribution_strength", 0.1)
self.include_indirect = self.config.get("include_indirect", True)
self.trace_residual = self.config.get("trace_residual", True)
self.collapse_threshold = self.config.get("collapse_threshold", 0.05)
# Track traced attribution maps
self.attribution_history = []
# Initialize glyph mappings for attribution
self._init_attribution_glyphs()
logger.info(f"Attribution tracer initialized for model: {model.model_id}")
def _init_attribution_glyphs(self):
"""Initialize glyph mappings for attribution visualization."""
# Attribution strength glyphs
self.strength_glyphs = {
"very_strong": "๐", # Very strong attribution (0.8-1.0)
"strong": "๐", # Strong attribution (0.6-0.8)
"moderate": "๐งฉ", # Moderate attribution (0.4-0.6)
"weak": "๐ซ๏ธ", # Weak attribution (0.2-0.4)
"very_weak": "๐ค", # Very weak attribution (0.0-0.2)
}
# Attribution type glyphs
self.type_glyphs = {
AttributionType.DIRECT: "โฎ", # Direct attribution
AttributionType.INDIRECT: "โค", # Indirect attribution
AttributionType.RESIDUAL: "โ", # Residual attribution
AttributionType.MULTIHEAD: "โฅ", # Multihead attribution
AttributionType.NULL: "โ", # Null attribution
AttributionType.COMPOSITE: "โฌฅ", # Composite attribution
AttributionType.RECURSIVE: "โป", # Recursive attribution
AttributionType.EMERGENT: "โ", # Emergent attribution
}
# Pattern glyphs
self.pattern_glyphs = {
"attribution_chain": "๐", # Clean attribution chain
"attribution_fork": "๐", # Forking attribution
"attribution_loop": "๐", # Loop in attribution
"attribution_gap": "โ", # Gap in attribution
"attribution_cluster": "โท", # Cluster of attribution
"attribution_decay": "๐", # Attribution decay
"attribution_conflict": "โก", # Conflicting attribution
}
# Meta glyphs
self.meta_glyphs = {
"attribution_focus": "๐ฏ", # Attribution focal point
"uncertainty": "โ", # Uncertainty in attribution
"recursive_reference": "๐", # Recursive attribution reference
"collapse_point": "๐", # Attribution collapse point
}
def trace(
self,
prompt: str,
output: Optional[str] = None,
depth: Optional[int] = None,
include_confidence: bool = True,
visualize: bool = False
) -> AttributionMap:
"""
Trace attribution between prompt and output.
Parameters:
-----------
prompt : str
Input prompt
output : Optional[str]
Output to trace attribution for. If None, will generate output.
depth : Optional[int]
Depth of attribution tracing. If None, uses default.
include_confidence : bool
Whether to include confidence scores
visualize : bool
Whether to generate visualization
Returns:
--------
AttributionMap
Map of attribution between tokens
"""
trace_start = time.time()
depth = depth or self.trace_depth
logger.info(f"Tracing attribution with depth {depth}")
# Generate output if not provided
if output is None:
output = self.model.generate(prompt=prompt, max_tokens=800)
# Get token-level representations
prompt_tokens = self._tokenize(prompt)
output_tokens = self._tokenize(output)
# Create attribution map
attribution_map = AttributionMap(
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
links=[],
metadata={
"prompt": prompt,
"output": output,
"model_id": self.model.model_id,
"trace_depth": depth,
"timestamp": time.time()
}
)
# If direct API access is available, get attribution directly
if hasattr(self.model, "get_attribution"):
try:
logger.info("Getting attribution directly from model API")
api_attribution = self.model.get_attribution(
prompt=prompt,
output=output,
include_confidence=include_confidence
)
attribution_map = self._process_api_attribution(
api_attribution,
prompt_tokens,
output_tokens
)
logger.info("Successfully processed API attribution")
except Exception as e:
logger.warning(f"Failed to get attribution from API: {e}")
logger.info("Falling back to inference-based attribution")
attribution_map = self._infer_attribution(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
depth=depth
)
else:
# Fall back to inference-based attribution
logger.info("Using inference-based attribution")
attribution_map = self._infer_attribution(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
depth=depth
)
# Analyze token salience
attribution_map.token_salience = self._analyze_token_salience(
attribution_map.links
)
# Find attribution gaps
attribution_map.attribution_gaps = self._find_attribution_gaps(
attribution_map.links,
len(prompt_tokens),
len(output_tokens)
)
# Detect collapsed regions
attribution_map.collapsed_regions = self._detect_collapsed_regions(
attribution_map.links,
len(prompt_tokens),
len(output_tokens)
)
# Calculate uncertainty if requested
if include_confidence:
attribution_map.uncertainty = self._calculate_attribution_uncertainty(
attribution_map.links
)
# Generate visualization if requested
if visualize and self.visualizer:
visualization = self.visualizer.visualize_attribution(attribution_map)
attribution_map.metadata["visualization"] = visualization
# Record execution time
trace_time = time.time() - trace_start
attribution_map.metadata["trace_time"] = trace_time
# Add to history
self.attribution_history.append(attribution_map)
logger.info(f"Attribution tracing completed in {trace_time:.2f}s")
return attribution_map
def trace_with_forks(
self,
prompt: str,
output: Optional[str] = None,
fork_factor: int = 3,
include_confidence: bool = True,
visualize: bool = False
) -> Dict[str, Any]:
"""
Trace attribution with multiple fork paths.
Parameters:
-----------
prompt : str
Input prompt
output : Optional[str]
Output to trace attribution for. If None, will generate output.
fork_factor : int
Number of alternative attribution paths to generate
include_confidence : bool
Whether to include confidence scores
visualize : bool
Whether to generate visualization
Returns:
--------
Dict[str, Any]
Dictionary containing attribution map and fork paths
"""
trace_start = time.time()
logger.info(f"Tracing attribution with {fork_factor} fork paths")
# Generate output if not provided
if output is None:
output = self.model.generate(prompt=prompt, max_tokens=800)
# Get token-level representations
prompt_tokens = self._tokenize(prompt)
output_tokens = self._tokenize(output)
# Get base attribution map
base_attribution = self.trace(
prompt=prompt,
output=output,
include_confidence=include_confidence,
visualize=False
)
# Generate fork paths
fork_paths = []
# Identify conflict points for forking
conflict_points = self._identify_conflict_points(base_attribution)
# Generate alternative attribution paths
for i in range(fork_factor):
# Create fork with alternative attributions
fork_path = self._generate_fork_path(
base_attribution=base_attribution,
conflict_points=conflict_points,
fork_id=f"fork_{i+1}",
fork_index=i
)
fork_paths.append(fork_path)
# Create fork result
fork_result = {
"base_attribution": base_attribution,
"fork_paths": fork_paths,
"conflict_points": conflict_points,
"metadata": {
"prompt": prompt,
"output": output,
"model_id": self.model.model_id,
"fork_factor": fork_factor,
"timestamp": time.time()
}
}
# Generate visualization if requested
if visualize and self.visualizer:
visualization = self.visualizer.visualize_attribution_forks(fork_result)
fork_result["metadata"]["visualization"] = visualization
# Record execution time
trace_time = time.time() - trace_start
fork_result["metadata"]["trace_time"] = trace_time
logger.info(f"Attribution fork tracing completed in {trace_time:.2f}s")
return fork_result
def trace_attention_heads(
self,
prompt: str,
output: Optional[str] = None,
layer_range: Optional[Tuple[int, int]] = None,
head_threshold: float = 0.1,
visualize: bool = False
) -> Dict[str, Any]:
"""
Trace attribution through specific attention heads.
Parameters:
-----------
prompt : str
Input prompt
output : Optional[str]
Output to trace attention for. If None, will generate output.
layer_range : Optional[Tuple[int, int]]
Range of layers to analyze. If None, analyzes all layers.
head_threshold : float
Minimum attention strength to include head
visualize : bool
Whether to generate visualization
Returns:
--------
Dict[str, Any]
Dictionary containing attention head analysis
"""
trace_start = time.time()
# Generate output if not provided
if output is None:
output = self.model.generate(prompt=prompt, max_tokens=800)
# Get token-level representations
prompt_tokens = self._tokenize(prompt)
output_tokens = self._tokenize(output)
# Get model information
model_info = self.model.get_model_info()
num_layers = model_info.get("num_layers", 12)
num_heads = model_info.get("num_heads", 12)
# Determine layer range
if layer_range is None:
layer_range = (0, num_layers - 1)
else:
layer_range = (
max(0, layer_range[0]),
min(num_layers - 1, layer_range[1])
)
# Trace attention head behavior
attention_heads = []
# If direct API access is available, get attention patterns directly
if hasattr(self.model, "get_attention_patterns"):
try:
logger.info("Getting attention patterns directly from model API")
attention_patterns = self.model.get_attention_patterns(
prompt=prompt,
output=output,
layer_range=layer_range
)
attention_heads = self._process_api_attention(
attention_patterns,
prompt_tokens,
output_tokens,
layer_range,
head_threshold
)
logger.info("Successfully processed API attention patterns")
except Exception as e:
logger.warning(f"Failed to get attention patterns from API: {e}")
logger.info("Falling back to inference-based attention analysis")
attention_heads = self._infer_attention_behavior(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
layer_range=layer_range,
num_heads=num_heads,
head_threshold=head_threshold
)
else:
# Fall back to inference-based attention analysis
logger.info("Using inference-based attention analysis")
attention_heads = self._infer_attention_behavior(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
layer_range=layer_range,
num_heads=num_heads,
head_threshold=head_threshold
)
# Analyze attention patterns
head_patterns = self._analyze_attention_patterns(attention_heads)
# Create attention result
attention_result = {
"prompt_tokens": prompt_tokens,
"output_tokens": output_tokens,
"attention_heads": attention_heads,
"head_patterns": head_patterns,
"metadata": {
"prompt": prompt,
"output": output,
"model_id": self.model.model_id,
"layer_range": layer_range,
"head_threshold": head_threshold,
"timestamp": time.time()
}
}
# Generate visualization if requested
if visualize and self.visualizer:
visualization = self.visualizer.visualize_attention_heads(attention_result)
attention_result["metadata"]["visualization"] = visualization
# Record execution time
trace_time = time.time() - trace_start
attention_result["metadata"]["trace_time"] = trace_time
logger.info(f"Attention head tracing completed in {trace_time:.2f}s")
return attention_result
def trace_qk_alignment(
self,
prompt: str,
output: Optional[str] = None,
layer_indices: Optional[List[int]] = None,
visualize: bool = False
) -> Dict[str, Any]:
"""
Trace query-key alignment in attention mechanisms.
Parameters:
-----------
prompt : str
Input prompt
output : Optional[str]
Output to trace alignment for. If None, will generate output.
layer_indices : Optional[List[int]]
Specific layers to analyze. If None, analyzes representative layers.
visualize : bool
Whether to generate visualization
Returns:
--------
Dict[str, Any]
Dictionary containing QK alignment analysis
"""
trace_start = time.time()
# Generate output if not provided
if output is None:
output = self.model.generate(prompt=prompt, max_tokens=800)
# Get token-level representations
prompt_tokens = self._tokenize(prompt)
output_tokens = self._tokenize(output)
# Get model information
model_info = self.model.get_model_info()
num_layers = model_info.get("num_layers", 12)
# Determine layer indices
if layer_indices is None:
# Sample representative layers (early, middle, late)
layer_indices = [
0, # First layer
num_layers // 2, # Middle layer
num_layers - 1 # Last layer
]
# Trace QK alignment
qk_alignments = []
# If direct API access is available, get QK values directly
if hasattr(self.model, "get_qk_values"):
try:
logger.info("Getting QK values directly from model API")
qk_values = self.model.get_qk_values(
prompt=prompt,
output=output,
layer_indices=layer_indices
)
qk_alignments = self._process_api_qk_values(
qk_values,
prompt_tokens,
output_tokens,
layer_indices
)
logger.info("Successfully processed API QK values")
except Exception as e:
logger.warning(f"Failed to get QK values from API: {e}")
logger.info("Falling back to inference-based QK alignment")
qk_alignments = self._infer_qk_alignment(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
layer_indices=layer_indices
)
else:
# Fall back to inference-based QK alignment
logger.info("Using inference-based QK alignment")
qk_alignments = self._infer_qk_alignment(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
layer_indices=layer_indices
)
# Analyze QK patterns
qk_patterns = self._analyze_qk_patterns(qk_alignments)
# Create QK alignment result
qk_result = {
"prompt_tokens": prompt_tokens,
"output_tokens": output_tokens,
"qk_alignments": qk_alignments,
"qk_patterns": qk_patterns,
"metadata": {
"prompt": prompt,
"output": output,
"model_id": self.model.model_id,
"layer_indices": layer_indices,
"timestamp": time.time()
}
}
# Generate visualization if requested
if visualize and self.visualizer:
visualization = self.visualizer.visualize_qk_alignment(qk_result)
qk_result["metadata"]["visualization"] = visualization
# Record execution time
trace_time = time.time() - trace_start
qk_result["metadata"]["trace_time"] = trace_time
logger.info(f"QK alignment tracing completed in {trace_time:.2f}s")
return qk_result
def trace_ov_projection(
self,
prompt: str,
output: Optional[str] = None,
layer_indices: Optional[List[int]] = None,
visualize: bool = False
) -> Dict[str, Any]:
"""
Trace output-value projection in attention mechanisms.
Parameters:
-----------
prompt : str
Input prompt
output : Optional[str]
Output to trace projection for. If None, will generate output.
layer_indices : Optional[List[int]]
Specific layers to analyze. If None, analyzes representative layers.
visualize : bool
Whether to generate visualization
Returns:
--------
Dict[str, Any]
Dictionary containing OV projection analysis
"""
trace_start = time.time()
# Generate output if not provided
if output is None:
output = self.model.generate(prompt=prompt, max_tokens=800)
# Get token-level representations
prompt_tokens = self._tokenize(prompt)
output_tokens = self._tokenize(output)
# Get model information
model_info = self.model.get_model_info()
num_layers = model_info.get("num_layers", 12)
# Determine layer indices
if layer_indices is None:
# Sample representative layers (early, middle, late)
layer_indices = [
0, # First layer
num_layers // 2, # Middle layer
num_layers - 1 # Last layer
]
# Trace OV projection
ov_projections = []
# If direct API access is available, get OV values directly
if hasattr(self.model, "get_ov_values"):
try:
logger.info("Getting OV values directly from model API")
ov_values = self.model.get_ov_values(
prompt=prompt,
output=output,
layer_indices=layer_indices
)
ov_projections = self._process_api_ov_values(
ov_values,
prompt_tokens,
output_tokens,
layer_indices
)
logger.info("Successfully processed API OV values")
except Exception as e:
logger.warning(f"Failed to get OV values from API: {e}")
logger.info("Falling back to inference-based OV projection")
ov_projections = self._infer_ov_projection(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
layer_indices=layer_indices
)
else:
# Fall back to inference-based OV projection
logger.info("Using inference-based OV projection")
ov_projections = self._infer_ov_projection(
prompt=prompt,
output=output,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
layer_indices=layer_indices
)
# Analyze OV patterns
ov_patterns = self._analyze_ov_patterns(ov_projections)
# Create OV projection result
ov_result = {
"prompt_tokens": prompt_tokens,
"output_tokens": output_tokens,
"ov_projections": ov_projections,
"ov_patterns": ov_patterns,
"metadata": {
"prompt": prompt,
"output": output,
"model_id": self.model.model_id,
"layer_indices": layer_indices,
"timestamp": time.time()
}
}
# Generate visualization if requested
if visualize and self.visualizer:
visualization = self.visualizer.visualize_ov_projection(ov_result)
ov_result["metadata"]["visualization"] = visualization
# Record execution time
trace_time = time.time() - trace_start
ov_result["metadata"]["trace_time"] = trace_time
logger.info(f"OV projection tracing completed in {trace_time:.2f}s")
return ov_result
def compare_attribution(
self,
prompt: str,
outputs: List[str],
include_confidence: bool = True,
visualize: bool = False
) -> Dict[str, Any]:
"""
Compare attribution across multiple outputs for the same prompt.
Parameters:
-----------
prompt : str
Input prompt
outputs : List[str]
List of outputs to compare attribution for
include_confidence : bool
Whether to include confidence scores
visualize : bool
Whether to generate visualization
Returns:
--------
Dict[str, Any]
Dictionary containing attribution comparison
"""
trace_start = time.time()
logger.info(f"Comparing attribution across {len(outputs)} outputs")
# Get token-level representation of prompt
prompt_tokens = self._tokenize(prompt)
# Trace attribution for each output
attribution_maps = []
for i, output in enumerate(outputs):
logger.info(f"Tracing attribution for output {i+1}/{len(outputs)}")
attribution_map = self.trace(
prompt=prompt,
output=output,
include_confidence=include_confidence,
visualize=False
)
attribution_maps.append(attribution_map)
# Compare attribution patterns
comparison = self._compare_attribution_maps(attribution_maps)
# Create comparison result
comparison_result = {
"prompt": prompt,
"prompt_tokens": prompt_tokens,
"outputs": outputs,
"attribution_maps": attribution_maps,
"comparison": comparison,
"metadata": {
"model_id": self.model.model_id,
"num_outputs": len(outputs),
"timestamp": time.time()
}
}
# Generate visualization if requested
if visualize and self.visualizer:
visualization = self.visualizer.visualize_attribution_comparison(comparison_result)
comparison_result["metadata"]["visualization"] = visualization
# Record execution time
trace_time = time.time() - trace_start
comparison_result["metadata"]["trace_time"] = trace_time
logger.info(f"Attribution comparison completed in {trace_time:.2f}s")
return comparison_result
def visualize_attribution(
self,
attribution_map: AttributionMap,
visualization_type: str = "network",
highlight_tokens: Optional[List[str]] = None,
output_path: Optional[str] = None
) -> Dict[str, Any]:
"""
Visualize attribution map.
Parameters:
-----------
attribution_map : AttributionMap
Attribution map to visualize
visualization_type : str
Type of visualization to generate
highlight_tokens : Optional[List[str]]
Tokens to highlight in visualization
output_path : Optional[str]
Path to save visualization to
Returns:
--------
Dict[str, Any]
Visualization result
"""
if self.visualizer:
return self.visualizer.visualize_attribution(
attribution_map=attribution_map,
visualization_type=visualization_type,
highlight_tokens=highlight_tokens,
output_path=output_path
)
else:
# Simple matplotlib visualization if no visualizer
return self._simple_visualization(
attribution_map=attribution_map,
visualization_type=visualization_type,
highlight_tokens=highlight_tokens,
output_path=output_path
)
def _tokenize(self, text: str) -> List[str]:
"""Tokenize text using model tokenizer."""
if hasattr(self.model, "tokenize"):
return self.model.tokenize(text)
else:
# Simple whitespace tokenization fallback
return text.split()
def _infer_attribution(
self,
prompt: str,
output: str,
prompt_tokens:
|